W4ashabii commited on
Commit
2732185
·
verified ·
1 Parent(s): b37cfd2

Add health v2 (CC BY 4.0) replacing v1

Browse files
README.md CHANGED
@@ -6,7 +6,7 @@ task_categories:
6
  - text-retrieval
7
  task_ids:
8
  - document-retrieval
9
- pretty_name: Nepali Embedding Bench (legal, gov)
10
  size_categories:
11
  - n<1K
12
  tags:
@@ -21,28 +21,33 @@ configs:
21
  data_files:
22
  - split: legal
23
  path: data/legal-*
 
 
24
  - split: gov
25
  path: data/gov-*
26
  - config_name: corpus
27
  data_files:
28
  - split: legal
29
  path: corpus/legal-*
 
 
30
  - split: gov
31
  path: corpus/gov-*
32
  ---
33
 
34
  # Nepali Embedding Bench
35
 
36
- Nepali (Devanagari) retrieval sets for **evaluating** embedding and retrieval models: legal and Government of Nepal, with a health split being rebuilt. Each set is a query file, a corpus of candidate passages and binary relevance judgements. The sets are **test only**. Please do not train on them.
37
 
38
  | split | queries | passages | author | license |
39
  |---|---|---|---|---|
40
  | `legal` | 179 | 1,074 | Siddhant Bhattarai | CC BY 4.0 |
 
41
  | `gov` | 171 | 1,026 | Siddhant Bhattarai | CC BY 4.0 |
42
 
43
  ## Layout
44
 
45
- There are two subsets, and each has a `legal` and a `gov` split.
46
 
47
  **`default`** has one row per query, with the passages that go with it:
48
 
@@ -51,7 +56,7 @@ There are two subsets, and each has a `legal` and a `gov` split.
51
  | `qid` | Query id |
52
  | `query` | The question |
53
  | `domain`, `subdomain` | Topic labels |
54
- | `positives` | Passages that answer the query, as a list of `{pid, title, text}`. Usually one. 6 legal and 11 gov queries have more than one: passages written as negatives for other queries that turned out to answer these. |
55
  | `negatives` | Five hard negatives as `{pid, title, text}`. Each is on the same topic as the query but does not answer it. |
56
 
57
  **`corpus`** pools every positive and negative in a split (`pid`, `title`, `text`). Use it to rank each query against the whole split. Scoring each query only against its own six passages is much easier: picking one of six at random already gets about 17% top-1.
@@ -76,6 +81,7 @@ Every flagged candidate was judged, and any that actually answers its query was
76
  | split | candidates judged | extra positives |
77
  |---|---|---|
78
  | legal | 375 | 8 |
 
79
  | gov | 76 | 11 |
80
 
81
  For legal and gov, two independent judging passes were reconciled, and disagreements got a second look. No negative states a conflicting value for the fact its query asks about. The candidate lists and the confirmed answers are in `reports/<split>_falseneg_candidates.jsonl` and `reports/<split>_falseneg_judged.jsonl`.
@@ -93,9 +99,15 @@ Source: [W4ashabii/legal_QA](https://huggingface.co/datasets/W4ashabii/legal_QA)
93
  - **21 items dropped**, so the ids have gaps. 18 ask a question that also appears in a training set (jiban4 or the gov-queries synthetic queries). 3 duplicate another question in the split: `legal_116`, `legal_164` and `legal_191` repeat `legal_042`, `legal_126` and `legal_147`.
94
  - **27 passages reworded** in plain language because their statute wording also appears in the Nepal gov web corpus. Each still states the same provision.
95
 
96
- ### health (being rebuilt)
97
 
98
- The previous health split has been removed. Its passages were Nepali translations of WHO fact sheets (CC BY-NC-SA 3.0 IGO and WHO website terms), NCBI StatPearls articles (CC BY-NC-ND 4.0) and Harvard Health pages (all rights reserved). None of these allow open reuse. A new health split, written fresh and released under CC BY 4.0, will replace it.
 
 
 
 
 
 
99
 
100
  ### gov
101
 
 
6
  - text-retrieval
7
  task_ids:
8
  - document-retrieval
9
+ pretty_name: Nepali Embedding Bench (legal, health, gov)
10
  size_categories:
11
  - n<1K
12
  tags:
 
21
  data_files:
22
  - split: legal
23
  path: data/legal-*
24
+ - split: health
25
+ path: data/health-*
26
  - split: gov
27
  path: data/gov-*
28
  - config_name: corpus
29
  data_files:
30
  - split: legal
31
  path: corpus/legal-*
32
+ - split: health
33
+ path: corpus/health-*
34
  - split: gov
35
  path: corpus/gov-*
36
  ---
37
 
38
  # Nepali Embedding Bench
39
 
40
+ Three Nepali (Devanagari) retrieval sets for **evaluating** embedding and retrieval models: legal, health and Government of Nepal. Each set is a query file, a corpus of candidate passages and binary relevance judgements. The sets are **test only**. Please do not train on them.
41
 
42
  | split | queries | passages | author | license |
43
  |---|---|---|---|---|
44
  | `legal` | 179 | 1,074 | Siddhant Bhattarai | CC BY 4.0 |
45
+ | `health` | 178 | 1,068 | Siddhant Bhattarai | CC BY 4.0 |
46
  | `gov` | 171 | 1,026 | Siddhant Bhattarai | CC BY 4.0 |
47
 
48
  ## Layout
49
 
50
+ There are two subsets, and each has a `legal`, a `health` and a `gov` split.
51
 
52
  **`default`** has one row per query, with the passages that go with it:
53
 
 
56
  | `qid` | Query id |
57
  | `query` | The question |
58
  | `domain`, `subdomain` | Topic labels |
59
+ | `positives` | Passages that answer the query, as a list of `{pid, title, text}`. Usually one. 6 legal, 6 health and 11 gov queries have more than one: passages written as negatives for other queries that turned out to answer these. |
60
  | `negatives` | Five hard negatives as `{pid, title, text}`. Each is on the same topic as the query but does not answer it. |
61
 
62
  **`corpus`** pools every positive and negative in a split (`pid`, `title`, `text`). Use it to rank each query against the whole split. Scoring each query only against its own six passages is much easier: picking one of six at random already gets about 17% top-1.
 
81
  | split | candidates judged | extra positives |
82
  |---|---|---|
83
  | legal | 375 | 8 |
84
+ | health | 113 | 7 |
85
  | gov | 76 | 11 |
86
 
87
  For legal and gov, two independent judging passes were reconciled, and disagreements got a second look. No negative states a conflicting value for the fact its query asks about. The candidate lists and the confirmed answers are in `reports/<split>_falseneg_candidates.jsonl` and `reports/<split>_falseneg_judged.jsonl`.
 
99
  - **21 items dropped**, so the ids have gaps. 18 ask a question that also appears in a training set (jiban4 or the gov-queries synthetic queries). 3 duplicate another question in the split: `legal_116`, `legal_164` and `legal_191` repeat `legal_042`, `legal_126` and `legal_147`.
100
  - **27 passages reworded** in plain language because their statute wording also appears in the Nepal gov web corpus. Each still states the same provision.
101
 
102
+ ### health
103
 
104
+ Health questions from the public in Nepal, each paired with a passage that answers it and five same-topic negatives. Passages are 35–60 words, in the style of a health-post leaflet or FAQ, and every passage in this split was written for it.
105
+
106
+ Domains: infectious disease (30), maternal and reproductive health (30), non-communicable disease (30), child health including Nepal's immunization schedule (29), mental health (15), health services (15), first aid and injury (15), nutrition and lifestyle (14).
107
+
108
+ Two items were dropped because they repeated another question in the split. Figures that change over time, such as insurance premiums, helpline numbers and programme details, are dated to 2026. This split is general health information, not medical advice.
109
+
110
+ **Version history.** The first health split (v1) was contributed by Sabin Bhattarai ([sabin1234/REAL_WEBSITE_TEXT_SCRAP_HEALTH_DOMAIN_DATASET](https://huggingface.co/datasets/sabin1234/REAL_WEBSITE_TEXT_SCRAP_HEALTH_DOMAIN_DATASET)). It was replaced for licensing reasons: its passages were Nepali translations of WHO fact sheets (CC BY-NC-SA 3.0 IGO and WHO website terms), NCBI StatPearls articles (CC BY-NC-ND 4.0) and Harvard Health pages (all rights reserved), and none of these allow open reuse. The current split (v2) reuses no text, questions or passages from v1. It is written by Siddhant Bhattarai and released under CC BY 4.0.
111
 
112
  ### gov
113
 
corpus/health-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3317be6b06876ed0533b1c00d87f60126a26f42058b58bb6ad616bc8dddd896
3
+ size 334976
data/health-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:baaa96e48827433ffc963e718ec1a45fc40031c21afa749bbbf2f3ac75d4c04c
3
+ size 337624
reports/health_drops.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "health_q079": "duplicate of health_q014 within the split",
3
+ "health_q143": "duplicate of health_q114 within the split"
4
+ }
reports/health_falseneg_candidates.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
reports/health_falseneg_judged.jsonl ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {"qid": "health_q011", "pid": "health_p0817", "candidate": "टाइफाइड साल्मोनेला टाइफी ब्याक्टेरियाले हुन्छ र यो सङ्क्रमित व्यक्तिको दिसा वा पिसाबले दूषित भएको खानेकुरा तथा पानीबाट सर्छ। सफा खानेपानी, उमालेको वा शुद्ध पानी पिउने र खाना बनाउनु अघि हात धुने बानीले यसको जोखिम घटाउँछ।", "reason": "Same content: Salmonella typhi, contaminated food/water spread"}
2
+ {"qid": "health_q013", "pid": "health_p0841", "candidate": "हैजा भिब्रियो कोलेरी ब्याक्टेरियाले हुन्छ, जसमा चामलको पानीजस्तो धेरै पातलो दिसा हुन्छ र शरीरमा चाँडै पानी घट्छ। यो दूषित पानी वा खानाबाट फैलिन्छ। प्रकोपको समयमा उमालेको पानी पिउने र चाँडो उपचार खोज्ने कुरा महत्त्वपूर्ण हुन्छ।", "reason": "States Vibrio cholerae causes cholera via contaminated water/food"}
3
+ {"qid": "health_q069", "pid": "health_p0579", "candidate": "आमाको पहिलो दूध (ब्रेस्ट मिल्क) मा बच्चालाई चाहिने धेरै पोषक तत्व र रोगसँग लड्ने पदार्थ हुन्छन्। जन्मदेखि छ महिनासम्म आमाको दूध मात्र खुवाउँदा बच्चाको स्वास्थ्य राम्रो रहन्छ र संक्रमण कम हुन्छ। यसले आमा र बच्चा दुवैलाई फाइदा पुर्‍याउँछ र परिवारको खर्च पनि घटाउँछ।", "reason": "States exclusive breastfeeding for six months from birth"}
4
+ {"qid": "health_q081", "pid": "health_p1038", "candidate": "झाडापखाला लाग्दा शिशुलाई आमाको दूध खुवाउन रोक्नु हुँदैन, बरु बढी पटक खुवाउनुपर्छ। छ महिनाभन्दा माथिका बच्चालाई पनि सजिलै पच्ने नरम खाना दिन जारी राख्नुपर्छ। खाना रोक्दा कमजोरी बढ्न सक्छ। राम्रो पोषणले आन्द्राको क्षतिपूर्ति छिटो गर्छ र बच्चालाई कुपोषणबाट जोगाउँछ।", "reason": "Do not stop breastfeeding in diarrhea; feed more often"}
5
+ {"qid": "health_q091", "pid": "health_p0125", "candidate": "स्वास्थ्यका लागि वयस्कले दिनमा करिब ५ ग्राम (एक चम्चाभन्दा कम) भन्दा बढी नुन खानु हुँदैन। धेरै नुनले रक्तचाप बढाउँछ र मुटु तथा मिर्गौलाको रोगको जोखिम बढाउँछ। अचार, नमकिन र प्रशोधित खानामा पनि नुन धेरै हुन्छ।", "reason": "Adults max about 5 g salt daily; raises blood pressure"}
6
+ {"qid": "health_q104", "pid": "health_p0180", "candidate": "सुर्ती, खैनी, गुटखा, पान र रक्सी सेवन गर्नेहरूमा मुखको क्यान्सरको जोखिम बढी हुन्छ। मुखमा निको नहुने घाउ, सेतो वा रातो दाग वा गाँठो देखिए ढिलो नगरी जाँच गराउनुपर्छ। सुरुमै पत्ता लगाउँदा उपचार सजिलो हुन्छ। त्यसैले सुर्तीजन्य पदार्थ र रक्सीबाट टाढा बस्नु मुखको स्वास्थ्यका लागि उपयोगी हुन्छ।", "reason": "Khaini, gutkha, tobacco, alcohol raise oral cancer risk"}
7
+ {"qid": "health_q104", "pid": "health_p0759", "candidate": "खैनी, गुटखा, पान मसला र जर्दा जस्ता धुवाँरहित सुर्तीजन्य पदार्थ पनि सुरक्षित होइनन्। यिनले मुख, घाँटी र खाने नलीको क्य��न्सर, मुखको घाउ र लत निम्त्याउन सक्छन्। धुम्रपानभन्दा सुरक्षित भन्ने धारणा गलत हो। यस्ता पदार्थ किशोरकिशोरीमा सजिलै पुग्ने भएकाले यसको रोकथाम गर्न जनचेतना चाहिन्छ।", "reason": "Khaini and gutkha cause oral cancer, not safe"}
reports/health_overlap_report.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checked_against": [
3
+ "gov",
4
+ "jiban4",
5
+ "legal_qa"
6
+ ],
7
+ "method": "NFC + punctuation/digit normalisation; exact, containment, word 5-gram containment vs single reference text (items with < 8 grams: exact/containment only)",
8
+ "flag_threshold": 0.5,
9
+ "review_threshold": 0.2,
10
+ "passages": {
11
+ "n": 1080,
12
+ "flagged": {},
13
+ "review": {}
14
+ },
15
+ "queries": {
16
+ "n": 180,
17
+ "flagged": {},
18
+ "review": {}
19
+ }
20
+ }
reports/health_semantic_report.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "BAAI/bge-m3",
3
+ "checked_against": [
4
+ "gov",
5
+ "jiban4",
6
+ "legal_qa"
7
+ ],
8
+ "n_reference_questions": 11836,
9
+ "flag_threshold": 0.85,
10
+ "review_threshold": 0.78,
11
+ "flagged": {},
12
+ "review": {
13
+ "health_q123": {
14
+ "score": 0.816,
15
+ "query": "आत्महत्या गर्न सक्ने व्यक्तिमा देखिने चेतावनीका संकेतहरू के के हुन्?",
16
+ "source": "jiban4",
17
+ "ref_id": "row:5949",
18
+ "ref_query": "आत्महत्याका लागि केही जोखिम कारकहरू के हुन्?"
19
+ },
20
+ "health_q153": {
21
+ "score": 0.8,
22
+ "query": "संविधानको कुन धारामा निःशुल्क आधारभूत स्वास्थ्य सेवा पाउने हकको व्यवस्था छ?",
23
+ "source": "jiban4",
24
+ "ref_id": "row:7792",
25
+ "ref_query": "नेपालको संविधान अनुसार स्वास्थ्यको हक के हो?"
26
+ },
27
+ "health_q162": {
28
+ "score": 0.792,
29
+ "query": "नेपालमा मानव अंग प्रत्यारोपण सम्बन्धी ऐन कुन सालमा बनेको हो?",
30
+ "source": "jiban4",
31
+ "ref_id": "row:5318",
32
+ "ref_query": "मानव अंग प्रत्यारोपण सम्बन्धी प्रावधानको उद्देश्य के हो?"
33
+ }
34
+ }
35
+ }